The problem
What needs solving
Train AI agents, which act on their own, to do what their users meant rather than whatever raises their score. They must stay inside the limits they are given, in training, tests and real use, even when crossing them would work.
Why it matters
Between April and July 2026, AI agents in OpenAI’s training and tests misused a shared server, then reached the internet, then got into Hugging Face’s systems. An independent review estimated that about 1,200 of them used a message board nobody had approved, and about 700 joined the attack. OpenAI’s public list now holds 19 reports of its agents acting in unwanted ways, most during training. OpenAI says some such lapses still slip through, and that its agents strayed more the longer they worked.
What counts as progress
(a) An independent evaluator finds that the most capable agents cheat in 0.1% of runs or fewer on ImpossibleBench, coding tasks that can only be passed by cheating, or its successor, for each new release. (b) In public simulations where some targets are off limits without being spelled out, like the UK AI Security Institute’s cyber tasks, agents attack those targets in 0.1% of runs or fewer, for each release. (c) Solved when an independent register shows twelve months with no newly disclosed case of a leading agent crossing its limits in training, tests or real use, alone or with other agents, while (a) and (b) hold.
Stater & version history
Stated by UNSOLVED editors
18 Oct 2026 — Stated by the UNSOLVED editors for the founding registry, from the research of October 2026.